OpenAI launches Astra for Law with GPT-6 Astra via Trusted Access
OpenAI's Astra for Law combines GPT-6 Astra with a legal search index and integrations for legal work. It is initially available to selected firms in ChatGPT and Codex, with API access planned later.

TL;DR
- Astra for Law packages GPT-6 Astra with a legal retrieval layer and specialized work instructions, and OpenAI's launch thread says selected firms will get it first in ChatGPT and Codex.
- Legal-research correctness rose from 38.7% to 54.0%, +15.3 points, in OpenAI's product evaluation.
- ChatGPT's legal catalog starts with 26 partner-built and 47 community plugins, according to gdb's overview.
- The published comparison is still provisional: ValsAI's validation update calls the reported runs OpenAI self-reported and says its team is validating them on a held-out set.
- Early build claims cover contract review, deal investigation, and IPO drafting, as reach_vb's launch note reports.
OpenAI's announcement already names the future API slug, gpt-6-astra-law, although that API is not public. A published Vals benchmark harness gives its agent web search, HTML parsing, and document retrieval tools, while Legal IT Insider's launch report names DeepJudge and iManage among the integration partners.
What shipped
- Astra for Law combines GPT-6 Astra, instructions for legal analysis and writing, and a dedicated Legal Search Index, as OpenAI's product description puts it.
- The index searches more than 230 million URLs across U.S. case law, statutes, regulations, court rules, and administrative decisions, with new sources added daily, per OpenAI's index description.
- OpenAI says the configuration is meant to give firms and legal technology companies an application layer they can adapt around their own workflows in its access announcement.
- The plugin catalog has 26 partner-built plugins and 47 community plugins, in gdb's overview of the launch.
Benchmarks that moved
First-party
- Overall correctness: 38.7% with GPT-6 Astra plus web search → 54.0% with Astra for Law, +15.3 points, in OpenAI's product evaluation.
- Weighted rubric score: 84.1% → 90.0%, +5.9 points, in the OpenAI self-reported runs that ValsAI's benchmark update describes.
- All-pass rate: 39.6% → 53.7%, +14.1 points, in those same ValsAI-reported runs.
Third-party evaluators
- ValsAI says it is validating the figures on its held-out test set in the evaluator's update.
- The published Vals harness evaluates tool-using agents on U.S. statutes, regulations, and case law across federal and state jurisdictions.
Customer-reported
- OpenAI's partner thread describes three co-developed products, but supplies no customer-reported performance metric.
Where it regressed
- General availability: OpenAI's access announcement limits the initial release to selected firms through Trusted Access in ChatGPT and Codex, with API access described as coming soon.
- Corpus scope: OpenAI's index description defines the Legal Search Index around U.S. legal materials.
- Price transparency: the OrcaRouter launch analysis reports that OpenAI has not published pricing for the legal configuration.
- Independent confirmation: the head-to-head figures remain OpenAI self-reported while ValsAI validates the held-out set.
Under the hood
- The index draws on Free Law Project's CourtListener collection, which OpenAI's announcement says covers roughly 99.9% of published U.S. precedential case law, alongside licensed content that includes Thomson Reuters material.
- Its legal instructions cover applying authorities to client facts, developing arguments, analyzing deal terms and risk allocation, and separating holdings from dicta, according to OpenAI's product details.
- OpenAI says the index is intended to move a lawyer from matter facts to authorities and supporting passages they can examine themselves in its Legal Search Index description.
- The product adds settings for thorough work, and OpenAI says it will maintain the legal configuration while builders create applications and workflows in its launch thread.
Trusted Access
The ChatGPT model picker will call the configuration “GPT-6 Astra Law,” while OpenAI's announcement identifies the planned API model as gpt-6-astra-law.
For eligible firms, OpenAI's announcement describes additional privacy and governance controls, including Zero Data Retention for API use and exclusion of the relevant ChatGPT Enterprise work from human review. The OrcaRouter launch analysis notes that neither a public API date nor a price accompanied the launch.
Where it shows up
- Sullivan & Cromwell co-built an agreement analyzer, according to OpenAI's partner thread.
- Ropes & Gray co-built an M&A diligence system, while Cooley co-built GO Public for IPO preparation, OpenAI's partner thread says.
- Partner plugins connect tools from Thomson Reuters, Harvey, Legora, and iManage, alongside community plugins that lawyers and legal engineers can adapt, per OpenAI's plugin announcement.
- Spellbook says it is deploying Astra for Law in its product in Scott Stevenson's post.